torch_staintools.hash package

Submodules

torch_staintools.hash.color module

torch_staintools.hash.color.od_angle_hash64(od: Tensor, mask: Tensor | None = None) → Tensor

Compute the 64bit hash from od angles.

Parameters:

  • od

  • mask – target masking area for tissue foreground. If None then all regions are considered tissue.

Returns:

the hash

Return type:

torch.Tensor

torch_staintools.hash.dhash module

torch_staintools.hash.dhash.od_dhash(od: Tensor, out_h: int = 8, out_w: int = 8) → tuple[Tensor, Tensor]

torch_staintools.hash.hash_util module

torch_staintools.hash.hash_util.pack_bits_u64(bits: Tensor) → Tensor

bits * weight may cause overflow in some rare case.

Parameters:

bits

Returns:

B-dimensional uint64

Return type:

torch.Tensor

torch_staintools.hash.key module

torch_staintools.hash.key.key_from_od(od: Tensor, mask: Tensor) → List[Tuple[int, int, bytes, int]]
torch_staintools.hash.key.lbp_code_to_bytes(lbp_code: Tensor) → List[bytes]
Parameters:

lbp_code – B x R x D. R as number of regions. D can be: 256: all 8-neighborhood encoding 59: uniform pattern 10: rotation invariance

Returns:

List of bytes. Each element is a bytes array corresponding to the hash of a data point in the batch.

torch_staintools.hash.key.to_uint(h: Tensor) → List[int]

torch_staintools.hash.lbp module

torch_staintools.hash.lbp.hist_lbp8_256_to_59(code_f: Tensor, lut: Tensor) → Tensor

Uniform pattern compression. 256 -> 59

Parameters:

  • code_f – flattened LBP codes. B x num_regions x 256.

  • lut – Look-up table. 256-D

Returns:

histogram. B x num_regions x 59 (uint16).

Return type:

torch.Tensor

torch_staintools.hash.lbp.od_lbp8_hash(od: Tensor, thumb_size: int = 32, grid_h: int = 1, grid_w: int = 1, use_riu: bool = True, levels: int | None = 16, use_sqrt: bool = True) → Tensor
Parameters:

  • od – Input batch image in optical density. BxCxHxW.

  • thumb_size – size of the code thumbnail.

  • grid_h – grid height. how many regions along H. (region height = thumb // grid_height)

  • grid_w – grid width. how many regions along W. (region width = thumb // grid_width)

  • use_riu – whether further cast to RIU (10bytes)

  • levels – number of quantization levels. If None, them bypass.

  • use_sqrt – whether to further compress the frequency by sqrt.

Returns:

the histogram. B x (grid_h*grid_w) * D

grid_h*grid_w as how many grids; D = 256 -> original LBP (8bits) D = 59 -> Uniform Pattern

Return type:

torch.Tensor

torch_staintools.hash.lbp.u2_hist59_to_riu2_hist10(hist59: Tensor, lut_riu: Tensor) → Tensor

map the u2 (uniform pattern, 59) to riu (rotation-invariance, 10)

Parameters:

  • hist59

  • lut_riu

Returns:

Module contents